Deactivate caching in fulltest pipeline #294
Conversation
| @@ -1,7 +1,7 @@ | |||
| # NOTES: | |||
| # - The map syntax used for matrix is flagged red but actually works | |||
| # - This runs everything in Python 3.10, 3.11 and 3.12 | |||
| # - No environments are cached due to space limit | |||
There was a problem hiding this comment.
the old reason is in principle still true so why delete?
There was a problem hiding this comment.
But that's not the primary reason why we don't want caching here, right? I thought we explicitly want to trigger to full installation pipeline etc (including newest package versions etc) in order to have a true e2e test? While the old reason might still be valid, it doesn't matter here. Or how would you even want me to express that? I guess not like "... to perform a full end-to-end test. Even if we didn't care about e2e, environments would still not be cached due to space limits." 😄
There was a problem hiding this comment.
no it was certainly one of the reasons and still is, if this action creates lots of caches it might delete caches form CI because theres a limit, just keep it and add your reason, no reason to delete
|
closing due to #298 |
Due to continuing serialization problems that were thought to be related with caching, #277 deactivated core test caching and #294 was prepared to do the same for the full test environment. This PR reactivates caching and instead refactors the class layout of `SKLearnClusteringRecommender` in an attempt to fix the root cause. Mysteriously, the top-level import of `sklearn.mixture.GaussianMixture` seems to cause trouble. While the reason is still unclear, turning it into a lazy import (which will also become handy later when making `scikit-learn` an optional dependency) seems to resolve the problem. On a side note: deactivating slots for the recommenders solves the problem as well, which suggests that the root cause could be related to classes not being properly garbage collected (since `attrs` needs to create new classes when slots are activated), which could also explain that `GaussianMixtureClusteringRecommender` seemed to have improperly overridden methods after deserialization (for example, the `__repr__` of a created Gaussian mixture recommender correctly pointed to its own class before serialization but to the `__repr__` of `SKLearnClusteringRecommender` after serialization – but weirdly only when executed in `tox`).
This PR removes the last remaining caching steps from our pipeline due to the mysterious interference with serialization. Can be reactivated once the root cause is identified.